Abstract: In recent years, dust detection methods based on image recognition have received full attention and development because they do not have installation and detection range limitations, but the real-time and accuracy of existing methods still need to be improved. In view of this, we propose a dust image detection method based on the improved YOLOv5 algorithm. Firstly, the existing YOLOv5 algorithm backbone network and Neck network were improved, and the original backbone network was replaced by GhostNet, a lightweight network, to reduce network parameters, and then three feature layers were output. Then, for the three feature layers of the backbone network output, the attention mechanism CA is applied to increase the network accuracy. Finally, ablation experiments and comparative experiments were designed to verify the effectiveness of the improved algorithm. The experimental results show that the mean Average Precision (mAP) of the improved algorithm can reach 92.11% and the detection speed reaches 37 frames persecond.